Cross-Layer Survivability in WDM-Based Networks

Transcription

1 Cross-Layer Survivability in WDM-Based Networks Kayi Lee, Member, IEEE, Eytan Modiano, Senior Member, IEEE, Hyang-Won Lee, Member, IEEE, Abstract In layered networks, a single failure at a lower layer may cause multiple failures in the upper layers. As a result, traditional schemes that protect against single failures may not be effective in cross-layer networks. In this paper, we introduce the problem of maximizing the connectivity of layered networks. We show that connectivity metrics in layered networks have significantly different meaning than their single-layer counterparts. Results that are fundamental to survivable single-layer network design, such as the Max-Flow Min-Cut theorem, are no longer applicable to the layered setting. We propose new metrics to measure connectivity in layered networks and analyze their properties. We use one of the metrics, Min Cross Layer Cut, as the objective for the survivable lightpath routing problem, and develop several algorithms to produce lightpath routings with high survivability. This allows the resulting cross-layer architecture to be resilient to failures between layers. Index Terms Cross-layer survivability, max-flow, min-cut, connectivity, disjoint paths, survivable path set, lightpath routing, multi-commodity flow I. INTRODUCTION Modern communication networks are constructed using a layered approach, as shown in Figure. Such a network typically consists of an electronic packet switched network (such as IP); often this packet-switched network is built on top of one or more electronic circuit switched transport networks (e.g., ATM, SONET; sometimes neither or both); and these in turn are built upon a fiber network. This multitude of layers is used in order to simplify network design and operations. However, this layering also leads to certain inefficiencies and interoperability issues. In this paper, we focus on the impact of layering on network survivability. We examine this problem in the context of Wavelength Division Multiplexing (WDM) based networks; although the concepts discussed are equally applicable to other layered architectures (e.g., IP over ATM, ATM over SONET, etc.). In a WDM-based network, the logical topology is defined by a set of nodes and lightpaths connecting the nodes, while the physical topology is defined by a (possibly different) set of nodes and the fibers connecting them. For example, an IP-over-WDM network consists of IP routers that are connected using (WDM) lightpaths as shown in Figure. Each lightpath is realized by setting up a physical connection using one of the wavelength channels in the optical fibers. For networks where wavelength conversion capability [] is unavailable, the lightpaths are also subject to the wavelength continuity constraint, which requires the lightpath to use the same wavelength channel along the physical route [2]. The authors are with the Massachusetts Institute of Technology, Cambridge, MA ( {kylee, modiano, hwlee}@mit.edu) This work was supported by NSF grants CNS and CNS and by DTRA grant number HDTRA This work was presented in part at the IEEE INFOCOM conference, April Fig.. An IP-over-WDM network where the IP routers are connected using optical lightpaths. The logical links (arrowed lines on top) are formed using lightpaths (arrowed lines at the bottom) that are routed on the physical fiber (thick gray lines at the bottom). In general, the logical and physical topologies are not the same. Networks often rely on the logical layer for providing protection and restoration services. However, even when the logical topology is designed to tolerate single logical link failures, once the logical topology is embedded on the physical topology, the logical topology may no longer be survivable to single physical (fiber) link failures. This is because each physical fiber link may carry multiple lightpaths. Hence, the failure of a single fiber link can lead to the failure of multiple links in the logical topology, which may subsequently leave the logical topology disconnected. As a simple illustrative example, consider the physical and logical topologies shown in Figures 2(a) and (b). The lightpaths in the logical topology are routed over the physical topology in two different ways in Figures 2(c) and (d). In Figure 2(c), a failure of physical fiber (, 5) would cause lightpaths (, 5) and (3, 5) to fail. Consequently, node 5 will be disconnected from other nodes in the logical topology. On the other hand, in Figure 2(d), the logical topology will remain connected even if one of the fibers fails. The above example demonstrates that in a multi-layer network, a physical link failure can result in multiple logical link failures, and that the routing of the logical links on the physical topology has a big impact on the connectivity of the multi-layer network. In contrast to the simplified example of Figure 2, real-life networks are highly intertwined and layered. However, due to the lack of general understanding of the issues in crosslayer survivability, most existing protection and restoration mechanisms are based on principles that are applicable only to single-layer network environments, and are subject to crosslayer issues as illustrated above. Nearly all previous literature in cross-layer survivability focused on single fiber failure, where the problem of interest generally falls into two broad categories: finding primary and protection routes for a single lightpath [3] [6], or routing multiple lightpaths jointly for a given logical topology [7] [6]. The problem studied in this paper belongs to the second

2 2 (a) Physical Topology (c) Unsurvivable Routing (b) Logical Topology (d) Survivable Routing Fig. 2. Different lightpath routings can affect survivability category, with the focus on designing lightpath routings that are tolerant against the maximum number of physical failures. Some of the concepts introduced in this paper are generalizations of the survivable lightpath routing concept, which was first introduced in [7]. The same paper also developed an ILP formulation for survivable routing of arbitrary logical topologies, which was subsequently improved in [0] and [5]. The problem of routing logical rings survivably on the physical network was studied in [7], [3], [4]. In [] the authors introduced the notion of piecewise survivable mapping and developed an algorithm to compute survivable routings based on piecewise survivable components. The same technique was extended to compute lightpath routings that are survivable against k failures, for a fixed value of k [7]. To the best of our knowledge, this is the first paper that formally studies classical survivability theory in the context of layered networks. We show that standard survivability metrics, such as the minimum cut and maximum disjoint paths, that have been widely used in characterizing the survivability properties of single-layer networks, lose much of their meaning in the context of cross-layer architecture. In particular, the Max- Flow Min-Cut Theorem, which constitutes the foundations of network survivability theory and provides the mathematcial justification of the aforementioned metrics, no longer holds in the cross-layer context. Such a fundamental difference suggests that many basic issues of cross-layer survivability are largely not understood. In Section II, we investigate some combinatorial properties of layered graphs related to network survivability, and highlight the key difference from their single-layer counterparts. In Section III, we specify the requirements for cross-layer survivability metrics, and propose two new metrics, Min Cross Layer Cut and Weighted Load Factor, that measure the connectivity of multi-layer networks. In Section IV, we consider the survivable lightpath routing problem using the Min Cross Layer Cut as the objective, and develop several lightpath routing algorithms based on the multi-commodity flow formulation in order to maximize the cross-layer connectivity of the network. In Section V, we present the simulation results for these algorithms, along with some empirical studies of the metrics introduced in Section III. II. FLOWS, CUTS AND PATH SETS IN LAYERED GRAPHS In this section, we will study connectivity structures such as flows, cuts and paths in multi-layer graphs from a theoretical standpoint, in order to develop insights into cross-layer survivability. We will highlight the key difference in combinatorial properties between multi-layer graphs and single-layer graphs. In particular, we will show that fundamental survivability results, such as the Max Flow Min Cut Theorem, are no longer applicable to multi-layer networks. Consequently, metrics such as connectivity have significantly different meanings in the cross-layer setting. Such fundamental differences make it much more challenging to design survivable multi-layer networks. A. Max Flow vs Min Cut For single-layer networks, the Max-Flow Min-Cut Theorem [8] states that the maximum amount of flow passing from the source s to the sink t always equals the minimum capacity that needs to be removed from the network so that no flow can pass from s to t. In addition, if all links have integral capacity, then there exists an integral maximum flow. This implies the maximum number of disjoint paths between s and t is the same as the minimum cut between the two nodes. Hence, the term connectivity between two nodes can be used unambiguously to refer to different measures such as maximum disjoint paths or minimum cut, and this makes it a natural choice as the standard metric for measuring network survivability. Because of its fundamental importance, we would like to investigate the Max-Flow Min-Cut relationship for multi-layer networks. We first generalize the definitions of Max Flow and Min Cut for layered networks: Definition : In a multi-layer network, the Max Flow between two nodes s and t in the logical topology is the maximum number of physically disjoint s t paths in the logical topology. The Min Cut between two nodes s and t in the logical topology is the minimum number of physical links that need to be removed in order to disconnect the two nodes in the logical topology. We model the physical topology as a network graph G P = (V P, E P ), where V P and E P are the nodes and links in the physical topology. The logical topology is modelled as G L = (V L, E L ), where V L V P. The lightpath routing is represented by a set of binary variables fij st, where a logical link (s, t) uses physical fiber (i, j) if and only if fij st =. Let P st be the set of all s t paths in the logical topology. For each path p P st, let L(p) be the set of physical links used { by the logical path p, that is, L(p) = (s,t) p (i, j) f st ij = }. Then the Max Flow and Min Cut between nodes s and t can be formulated mathematically as follows: MaxFlow st : Maximize p:(i,j) L(p) p P st f p, subject to: f p (i, j) E P () f p {0, } p P st

3 3 MinCut st : Minimize (i,j) L(p) y ij, subject to: y ij p P st (2) y ij {0, } (i, j) E P The variable f p in the formulation MaxFlow st indicates whether the path p is selected for the set of (s, t)-disjoint paths. Constraint () requires that no selected logical paths share a physical link. Similarly, in the formulation MinCut st, the variable y ij indicates whether the physical fiber (i, j) is selected for the minimum (s, t)-cut. Constraint (2) requires that all logical paths between s and t traverse some physical fiber (i, j) with y ij =. Note that the above formulations generalize the the Max Flow and Min Cut for single-layer networks. In particular, the formulations model the classical Max Flow and Min Cut of a graph G if both G P and G L are equal to G, and fij st = if and only if (s, t) = (i, j). Let MaxFlow st and MinCut st be the optimal values of the above Max Flow and Min Cut formulations. We also denote MaxFlowst R and MinCut R st to be the optimal values to the linear relaxations of above Max Flow and Min Cut formulations. The Max-Flow Min-Cut Theorem for single-layer networks can then be written as follows: MaxFlow st = MaxFlow R st = MinCutR st = MinCut st. The equality among these values has profound implications on survivable network design for single-layer networks. Because all these survivability measures converge to the same value, it can naturally be used as the standard survivability metric that is applicable to measuring both disjoint paths or minimum cut. Another consequence of this equality is that linear programs (which are polynomial time solvable) can be used to find the minimum cut and disjoint paths in the network. It is therefore interesting to see whether the same relationship holds for multi-layer networks. First, it is easy to verify that the linear relaxations for the formulations MaxFlow st and MinCut st maintain a primal-dual relationship, which, by Duality Theorem [9], implies that MaxFlow R st =MinCutR st. In addition, since any feasible solution to an integer program is also a feasible solution to the linear relaxation, we can establish the following relationship: Observation : MaxFlow st MaxFlow R st = MinCut R st MinCut st. Therefore, like single-layer networks, the maximum number of disjoint paths between two nodes cannot exceed the minimum cut between them in a multi-layer network. However, unlike the single-layer case, the values of MaxFlow st, MaxFlow R st and MinCut st are not always identical, as illustrated in the following example. In our examples throughout the section, we use a logical topology with two nodes s and t that are connected by multiple lightpaths. For simplicity of exposition, we omit the complete lightpath routing and only show the physical links that are shared by multiple lightpaths. Theorem states that this simplification can be made without loss of generality. Theorem : Let G L be a logical topology with two nodes s and t, connected by n lightpaths E L = {e, e 2,..., e n }, and let R = {R, R 2,...,R k } be a family of subsets of E L, where each R i 2, that captures the fiber-sharing relationship of the logical links. There exist a physical topology G P = (V P, E P ) and lightpath routing of G L over G P, such that: ) there are exactly k fibers in E P, denoted by F = {f, f 2,..., f k }, that are used by multiple lightpaths; 2) for each fiber f i F, the set of lightpaths using f i is R i. Proof: See Appendix A. Theorem implies that for a two-node logical topology, any arbritrary fiber-sharing relationship R can be realized by reconstructing a physical topology and lightpath routing. Therefore, in the following discussion, we can simplify our examples by only giving the fiber-sharing relationship of our two-node logical topology without showing the details of the lightpath routing. In Figure 3, the two nodes in the logical topology are connected by three lightpaths. The logical topology is embedded on the physical topology in such a way that each pair of lightpaths share a fiber. It is easy to see that no single fiber can disconnect the logical topology, and that any pair of fibers would. Hence, the value of MinCut st is 2 in this case. On the other hand, the value of MaxFlow st is only, because any two logical links share some physical fiber, so none of the paths in the logical network are physically disjoint. Finally, the value of MaxFlowst R is.5 because a flow of 0.5 can be routed on each of the lightpaths without violating the capacity constraints at the physical layer. Therefore, all three quantities are different in this example. We will study the integrality gaps for the formulations more carefully. Fig. 3. A logical topology with 3 links where each pair of links shares a fiber in the physical topology ) Integrality Gap for MaxFlow st : The above example can be generalized to show that the ratio between MaxFlow st and MaxFlowst R is O(n), where n is the number of paths between s and t. Consider an instance of lightpath routing where the two nodes in the logical network are connected by n logical links, and every pair of logical links share a separate fiber. In this case, the value of MaxFlow st will be, and the value of MaxFlowst R will be n 2, using the same arguments as above. Therefore, the ratio MaxFlowst is O(n). Note that this MaxFlowst R is an asymptotically tight bound since MaxFlow st and MaxFlowst R n for all lightpath routings. 2) Integrality Gap for MinCut st : The ratio between MinCut st and MinCut R st can be shown to be at most O(log n) as a direct application of the result by Lovász [20], who showed that the integrality gap between integral and fractional set cover is O(log n). We can construct a lightpath routing

4 4 where the gap between the two values is O(log n), thereby showing the tightness of the bound. Consider a layered network consisting of a two-node logical topology, and a set of k fibers F = {f,..., f k } that are shared by multiple logical links. For every subset T of k 2 + fibers in F, we add a logical link between the two logical nodes that uses only the fibers in T. Hence, for every set of k 2 fibers, there is a logical link that does not use any of the fibers. This implies the Min Cut is at least k 2. On the other hand, since each logical link uses exactly k 2 + fibers, the assignment where each y ij = satisfies k 2 + Constraint (2), and is therefore a feasible solution to MinCut R st. The objective value of this solution is k k 2 +, which is at most 2. Therefore, the integrality gap MinCutst is at least k MinCut R st 4. ( Therefore, ) for the two-node logical network with n = k logical links, the ratio between the integral and k 2 + relaxed optimal values for the Min Cut is O(k) = O(log n). We summarize our observation as follows: Observation 2: In a layered network, the values of MaxFlow st, MaxFlow R st and MinCut st can be all different. In addition, the gaps among the three values are not bounded by any constant. Therefore, a multi-layer network with high connectivity value (i.e. that tolerates a large number of failures) does not guarantee existence of physically disjoint paths. This is in sharp contrast to single-layer networks where the number of disjoint paths is always equal to the minimum cut. It is thus clear that network survivability metrics across layers are not trivial extensions of the single layer metrics. New metrics need to be carefully defined in order to measure crosslayer survivability in a meaningful manner. In Section III, we will specify the requirements for cross-layer survivability metrics, and propose two new metrics that can be used to measure the connectivity of multi-layer networks. B. Minimum Survivable Path Set In this section, we introduce another graph structure, called Survivable Path Set, that is useful in describing connectivity in layered networks. A survivable path set for two logical nodes s and t is the smallest set of s t logical paths such that at least one of the paths in the set survives for any single physical link failure. The Minimum Survivable Path Set, denoted as MinSPS st, is the size of the smallest survivable path set. For convenience, MinSPS st is defined to be if no survivable path set exists. In a single layer network, the value of MinSPS st reveals nothing more than the existence of disjoint paths, as its value is either two or, depending on whether disjoint paths between s and t exist. However, for multi-layer networks, MinSPS st can take on other values. For example, in Figure 3, the minimum survivable path set for s and t has size three because any pair of logical links can be disconnected by a single fiber failure. In fact, it is easy to verify that: MinSPS st = 2 if and only if MaxFlow st 2; MinSPS st = if and only if MinCut st =. Therefore, the value of MinSPS st provides a different perspective about the connectivity between two nodes in the cross-layer setting. It is particularly interesting in the regime where MaxFlow st = and MinCut st 2, i.e., there is a gap between the Max Flow and the Min Cut. The following theorem reveals a connection between survivable path sets and the relaxed maxflow MaxFlow st R. Theorem 2: MinSPS st log EP +. log MaxFlowst R Proof: See Appendix B It is worth noting that the theorem provides a sufficient condition for the existence of disjoint paths in the layered networks, in terms of the optimal value of MaxFlowst: R Corollary 3: Disjoint paths between two nodes s and t exist in a layered network if the relaxed Max Flow, MaxFlowst, R is greater than E P. Proof: By Theorem 2, a survivable path set of size two exists if MaxFlowst R > E P. This implies the existence of s t disjoint paths in the layered network. Therefore, survivable path sets not only are interesting graph structures that describe connectivity of layered networks, they can also be useful in revealing the relationship between integral and fractional flows in the layered network. C. Computational Complexity For single-layer networks, because the integral Max Flow and Min Cut values are always identical to the optimal relaxed solutions, these values can be computed in polynomial time [8]. However, computing and approximating their crosslayer equivalents turns out to be much more difficult. Theorem 4 describes the complexity of computing the Max Flow and Min Cut for multi-layer networks. Theorem 4: Computing Max Flow and Min Cut for multilayer networks is NP-hard. In addition, both values cannot be approximated within any constant factor, unless P=NP. Proof: The Max Flow can be reduced from the NP-hard Maximum Set Packing problem [2]: Maximum Set Packing: Given a set of elements E = {e, e 2,..., e n } and a family F = {C, C 2,...,C m } of subsets of E, find the maximum value k such that there exist k subsets {C j, C j2,...,c jk } F that are mutually disjoint. Given an instance of Maximum Set Packing, we construct a 2-node logical topology connected by multiple lightpaths as described in Theorem, so that the optimal value of the Maximum Set Packing instance equals the maximum number of physically disjoint paths in the 2-node logical topology. This means that Maximum Set Packing is polynomial time reducible to the 2-node disjoint path problem. Theorem implies that any instance of the 2-node disjoint path problem is polynomial time reducible to an instance of the multi-layer Max Flow problem. It follows that Maximum Set Packing is polynomial time reducible to the multi-layer Max Flow problem, which proves the theorem. Given an instance of Maximum Set Packing with ground set E and a family F of subsets of E, we construct a logical topology with two nodes, s and t, connected by F logical

5 5 links, where each logical link corresponds to a subset in F. The logical links are embedded on the physical network in a way that two logical links share a physical fiber if and only if their corresponding subsets share a common element in the Maximum Set Packing instance. It immediately follows that a set of physically disjoint s t paths in the logical topology corresponds to a family of mutually disjoint subsets of E. Similarly, the Min Cut can be reduced from the NP-hard Minimum Set Cover problem [2]: Minimum Set Cover: Given a set E = {e, e 2,..., e n } and a family F = {C, C 2,...,C m } of subsets of E, find the minimum value k such that there exist k subsets {C j, C j2,..., C jk } F that cover E, i.e., C jl C = E. Given an instance of Minimum Set Cover with ground set E and family of subsets F, we construct a logical topology that contains two nodes connected by a set of E logical links, where each logical link l i corresponds to the element e i. The logical links are embedded on the physical network in a way that exactly F fibers, namely { f,...,f F }, are used by multiple logical links, and the logical link l i uses physical fiber f j if and only if e i C j. It follows that the minimum number of physical fibers that forms a cut between the two logical nodes equals the size of a minimum set cover. The inapproximability result follows immediately from the inapproximabilities of the Maximum Set Packing and Minimum Set Cover problems [22] [24]. In summary, multi-layer connectivity exhibits fundamentally different structural properties from its single-layer counterpart. Because of that, it is important to reinvestigate issues of quantifying, measuring as well as optimizing survivability in multi-layer networks. In the rest of the paper, we will focus on designing appropriate metrics for layered networks, and developing algorithms to maximize the cross-layer survivability. III. METRICS FOR CROSS-LAYER SURVIVABILITY The previous section demonstrates the new challenges in designing survivable layered network architectures. Insights into quantifying and optimizing survivability are fundamentally different between the single-layer and multi-layer settings. In this section, we focus on the issue of quantifying survivability in multi-layer networks. Not only should such metrics have natural physical meaning in the cross-layer setting, they should also be mathematically consistent and compatible with the conventional single-layer connectivity metric. Hence, we first define formal requirements for metrics that can be used to quantify cross-layer survivability: Consistency: A network with a higher metric value should be more resilient to failures. Monotonicity: Any addition of physical or logical links to the network should not decrease the metric value. Compatibility: The metric should generalize the connectivity metric for single-layer networks. In particular, when applied to the degenerated case where the physical and logical topologies are identical, the metric should be equivalent to the connectivity of the topology. A metric that carries all the above properties would give us a meaningful and consistent measure of survivability for layered networks. We propose two metrics, the Min Cross Layer Cut and the Weighted Load Factor, that can be used as cross-layer survivability metrics. It is easy to verify that both metrics satisfy the above requirements. A. Min Cross Layer Cut In Section II, we defined MinCut st to be the minimum number of physical failures that would disconnect logical nodes s and t. One can easily generalize this by taking the minimum over all possible node pairs to obtain a global connectivity metric. We define the Min Cross Layer Cut (MCLC) to be the minimum number of physical failures that would disconnect the logical topology. A lightpath routing with high Min Cross Layer Cut value implies that the network remains connected even after a large number of physical failures. It is also a generalization of the survivable lightpath routing definition in [7], since a lightpath routing is survivable if and only if its Min Cross Layer Cut is greater than. Let S be a subset of the logical nodes V L, and δ(s) be the set of the logical links with exactly one end point in S. Let H S be the minimum number of physical links failures required to disconnect all links in δ(s). The Min Cross Layer Cut can be defined as follows: MCLC = min S V L H S. For each S, computing H S can be considered as finding the Min Cut between the two partitions S and V L S. In the proof of Theorem 4, we have shown that computing the value of MinCut st is NP-Hard even if the logical topology contains just two nodes. This immediately implies that computing the global MCLC value is NP-Hard: Theorem 5: Computing the MCLC is NP-Hard. In practice, however, the MCLC is bounded by the node degree of the logical topology, which is usually a small constant d. In that case, the MCLC can be computed in polynomial time by enumerating all fiber sets with up to d fibers. To compute the MCLC in a general setting, it can be modelled by the following integer linear program. Given the physical and logical topologies (V P, E P ), and (V L, E L ), let fij st be binary constants that represent the lightpath routing, such that logical link (s, t) uses physical fiber (i, j) if and only if fij st =. The MCLC can be formulated as the integer program below: M MCLC : Minimize y ij, subject to: d t d s y ij fij st (s, t) E L (3) d n, d 0 = 0 (4) n V L d n, y ij {0, } n V L, (i, j) E P The integer program contains a variable y ij for each physical link (i, j), and a variable d k for each logical node

6 6 k. Constraint (3) maintains the following property for any feasible solution: if d k =, the node k will be disconnected from node 0 after all physical links (i, j) with y ij = are removed. To see this, note that since d k = and d 0 = 0, any logical path from node 0 to node k contains a logical link (s, t) where d s = 0 and d t =. Constraint (3) requires that such a logical link traverse at least one of the fibers (i, j) with y ij =. As a result, all paths from node 0 to node k must traverse one of these fibers, and node k will be disconnected from node 0 if these fibers are removed from the network. Constraint (4) requires node 0 to be disconnected from at least one node, which ensures that the set of fibers (i, j) with y ij = forms a global Cross Layer Cut. In Section IV, we will use MCLC as the objective for the survivable lightpath routing problem, and develop algorithms to maximize such an objective. B. Weighted Load Factor Another way to measure the connectivity of a layered network is by quantifying the impact of each physical failure. The Weighted Load Factor (WLF), an extension of the metric Load Factor introduced in [25], provides such a measure of survivability. Given the physical topology (V P, E P ) and logical topology (V L, E L ), let fij st be binary constants that represent the lightpath routing, such that logical link (s, t) uses physical fiber (i, j) if and only if fij st =. We formulate the WLF as follows: z M WLF : Maximize w st w st > 0 z, subject to: w st fij st S V L, (i, j) E P S V L 0 z, w st (s, t) E L, where δ(s) is the cut set of S, i.e., the set of logical links that have exactly one end point in S. The variables w st are the weights assigned to the lightpaths. Over all possible logical cuts, the variable z measures the maximum fraction of weight inside a cut carried by a fiber. Intuitively, if we interpret the weight to be the amount of traffic in the lightpath, the value z can be interpreted as the maximum fraction of traffic across a set of nodes disrupted by a single fiber cut. The Weighted Load Factor formulation, defined to maximize the reciprocal of this fraction, thus tries to compute the logical edge weights that minimize the maximum fraction. This effectively measures the best way of spreading the weight across the fibers for the given lightpath routing. A lightpath routing with a larger Weighted Load Factor value means that it is more capable of spreading its weight within any cut across the fibers. The Weighted Load Factor also generalizes the survivable lightpath routing defined in [7], since its value will be greater than if and only if the lightpath routing is survivable. Although the formulation M WLF contains the quadratic terms zw st, the optimal value of z can be obtained by iteratively solving the linear program with different fixed values of z. Using binary search over the range of z, we can find the minimum z where a feasible solution exists. Unfortunately, the formulation M WLF contains an exponential number of constraints, and may not be polynomial time solvable. In fact, Theorem 6 states that finding objective value for M WLF is NP-Hard, even if the weights of the logical links w st are given. Theorem 6: Computing the Weighted Load Factor for a lightpath routing is NP-Hard even if the weight assignment w st for the logical links is fixed. Proof: The NP-Hardness proof is based on the reduction from the NP-Hard Uniform Sparsest Cut [26] problem. For details, see Appendix C. Finally, Theorem 7 describes the relationship between the WLF and the MCLC. Given a lightpath routing, let M MCLC be the ILP formation for its Min Cross Layer Cut, and let MCLC and MCLC R be the optimal values for M MCLC and its linear relaxation respectively. In addition, let W LF be the Weighted Load Factor of the lightpath routing. Then we have the following relationship: Theorem 7: MCLC R WLF MCLC. Proof: See Appendix D. Therefore, although the two metrics appear to measure different aspects of network connectivity, they are inherently related. As we will show in Section V, the two values are often identical. IV. LIGHTPATH ROUTING FOR MCLC MAXIMIZATION In this section, we consider the survivable lightpath routing problem using MCLC as the objective. At an abstract level, the optimal lightpath routing can be expressed as the following optimization problem: max f F min MFC(f, S), S V L where F is set of all possible lightpath routings, V L is the logical node set, and MFC(f, S) is the minimum number of fibers whose removal will disconnect all logical links in the cut set δ(s) given the lightpath routing f. This is a Max-Min- Min problem that may not have a simple formulation. In [27], we present an ILP formulation for this optimization problem. However, the formulation contains an exponential number of variables and constraints, which makes it infeasible even for small networks. Therefore, in this section, we consider ILP formulations whose objective values are lower bounds to the MCLC. These formulations are much simpler than the exact formulation presented in [27]. This makes it possible to develop survivable lightpath routing algorithms based on these simpler formulations. In particular, in Section IV-C we discuss how to use randomized rounding [28] based on these formulations as a heuristic to approximate MCLC maximization. Note that since MCLC is O(log n) inapproximable, polynomial time algorithms with approximation guarantees within this factor

7 7 are unlikely to exist. Therefore, we will instead evaluate the performance of our algorithms via simulation in Section V. All the formulations introduced in this section are based on multi-commodity flows, where each lightpath is considered a commodity to be routed over the physical network. Given the physical network G P = (V P, E P ) and the logical network G L = (V L, E L ), the multi-commodity flow for a lightpath routing can be generally formulated as follows: MCF X : Minimize X(f), f st ij {0, } subject to: {f st ij : (i, j) E P } forms an (s, t)-path, (s, t) E L, (5) where f is the variable set that represents the lightpath routing, such that fij st = if and only if lightpath (s, t) uses physical fiber (i, j) in its route. X(f) is an objective function that depends on f. For WDM networks where the wavelength continuity constraint is present, the above formulation can be extended to capture the wavelength assignment aspect. In that case, the wavelength assignment can be modelled by replacing the variable set fij st by fijλ st, which equals if and only if lightpath (s, t) uses wavelength λ on physical link (i, j). Constraint (5) can be{ easily extended } to restrict that, for each logical link (s, t), fijλ st = forms an (s, t) physical path along one of the wavelengths. To make sure that any wavelength λ on a physical fiber is used by at most one lightpath, the following constraint will be added: fijλ st (i, j) E P, λ. (6) (s,t) E L Similar formulations based on multi-commodity flows with wavelength continuity constraint have been proposed to solve the RWA problem of WDM networks [29], [30], where the objective is to minimize the number of lightpaths that traverse the same fiber. The key difference in the problem studied in this paper is in the objective function X, which should instead describe the survivability of the lightpath routing. To focus on the survivability aspect of the problem, the wavelength continuity constraint will be omitted in the formulations below. However, in cases where the wavelength continuity constraint is necessary, all these formulations can be extended as discussed above. A. Simple Multi-Commodity Flow Formulations Ideally, to ensure that the lightpath routing is survivable against the largest number of failures, the objective function X(f) should express the MCLC value of the lightpath routing given by f. However, since there is no simple way to express the lightpath routing problem that maximizes the MCLC as an integer linear program, we use an objective that approximates the MCLC value. In our formulation, each lightpath is assigned a weight w. The objective function measures the maximum load of the fibers, where the load is defined to be the total lightpath weight carried by the fiber. The intuition is that the multi-commodity flow formulation will try to spread the weight of the lightpaths across multiple fibers, thereby minimizing the impact of any single fiber failure. We can formulate an ILP with such an objective as follows: MCF w : Minimize, subject to: w(s, t)fij st (i, j) E P (s,t) E L f st ij {0, } {f st ij :(i, j) E P } forms an (s, t)-path, (s, t) E L As we will prove in Theorem 8, with a careful choice of the weight function w, the value gives a lower bound on the MCLC. Therefore, a lightpath routing with a low value is guaranteed to have a high MCLC. The routing strategy of the algorithm is determined by the weight function w. For example, if w is set to for all lightpaths, the integer program will minimize the number of lightpaths traversing the same fiber. Effectively, this will minimize the number of disconnected lightpaths in the case of a single fiber failure. In order to customize MCF w towards maximizing the MCLC of the solution, we propose a different weight function w MinCut that captures the connectivity structure of the logical topology. For each edge (s, t) E L, we define w MinCut (s, t) to be MinCut, where MinCut L(s,t) L(s, t) is the minimum (s, t)-cut in the logical topology. Therefore, if an edge (s, t) belongs to a smaller cut, it will be assigned a higher weight. The algorithm will therefore try to avoid putting these small cut edges on the same fiber. If w MinCut is used as the weight function used in MCF w, we can prove the following relationship between the objective value of a feasible solution to MCF w and the Weighted Load Factor of the associated lightpath routing: Theorem 8: For any feasible solution f of MCF w with w MinCut as the weight function, WLF. Proof: By definition of the weight function w MinCut, given any S V L, every edge in δ(s) has weight at least. Therefore, we have: δ(s) w(s, t) =. (7) δ(s) Now consider the lightpath routing associated with f. For any logical cut δ(s), the maximum fraction of weight inside the cut carried by a fiber is: max max max w(s, t)f st ij w(s, t) w(s, t)f st ij, by Equation (7) w(s, t)fij st. (s,t) E L In other words, no fiber in the network is carrying more than a fraction of the weight in any cut. This gives us a feasible solution to the Weighted Load Factor formulation M WLF, where each variable w st is assigned the value of

8 8 w MinCut (s, t), and the variable z is assigned the value of. As a result, the Weighted Load Factor, defined to be the maximum value of z among all feasible solutions to M WLF, must be at least. As a result of Theorems 7 and 8, the MCLC of a lightpath routing is lower bounded by the value of, which the algorithm will try to maximize. B. Enhanced Multi-Commodity Flow Formulation As we have discussed in Section III-B, the Weighted Load Factor provides a good lower bound on the MCLC of a lightpath routing. Here we propose another multi-commodity flow based formulation whose objective function approximates the Weighted Load Factor of a lightpath routing. The formulation, denoted as MCF LF, can be written as follows: MCF LF : Minimize γ, γ δ(s) f st ij subject to: (i, j) E P, S V L f st ij {0, } {f st ij :(i, j) E P } forms an (s, t)-path, (s, t) E L Essentially, the formulation optimizes the unweigthed Load Factor of the lightpath routing, (i.e., all weights equal one), by minimizing the maximum fraction of a logical cut carried by a single fiber. As this formulation provides a constraint for each logical cut, it captures the impact of a single fiber cut on the logical topology in much greater detail. The following theorem shows that for any lightpath routing, its associated Load Factor value γ gives a tighter lower bound than, given by the MCF w formulation. Theorem 9: For any lightpath routing, let be its associated objective value in the formulation MCF w with w MinCut as the weight function, and let γ be its associated objective value in the formulation MCF LF. In addition, let WLF be its Weighted Load Factor. Then: γ WLF. Proof: The value γ is the objective value for the formulation M WLF in Section III-B when all logical links have weight. This gives a feasible solution to M WLF, and implies that WLF γ. To prove that γ, we consider the physical link (i, j) and logical cut set δ(s) where (i, j) carries a fraction γ of the logical links in δ(s). Let L ij be the set of logical links in E L carried by (i, j). Therefore, we have γ = Lij δ(s) δ(s). In addition, by the definition of, we have w(s, t) δ(s) = γ. (s,t) L ij (s,t) L ij δ(s) This implies γ. Therefore, the formulation MCF LF gives a lightpath routing that is optimized for a better lower bound on the MCLC. However, this comes at the cost of a larger number of constraints and solving such an integer program may not be feasible in practice. Therefore, we next introduce a randomized rounding technique that approximates the optimal lightpath routing by solving the linear relaxation of the integer program. As we will see in Section V, the randomized rounding technique significantly speeds up the running time of the algorithm without observable degradation in the MCLC performance. This offers a practical alternative to solving the integer program formulations introduced in this section. C. Randomized Rounding for Lightpath Routing While the multi-commodity flow integer program formulations discussed in the previous section introduce a novel way to route lightpaths in a survivable manner, such an approach may not scale to large networks, due to the inherent complexity of solving integer programs. In order to circumvent the computational difficulty, we apply the randomized rounding technique, which is able to quickly obtain a near-optimal solution to the integer program. Randomized rounding has previously been used to solve multi-commodity flow problems to minimize the link load [28], [29], and its performance guarantee is studied in [28]. Through randomized rounding, we solve the linear relaxation of the ILP, and choose a random physical path between s and t for each lightpath (s, t) with probability based on the optimal fraction solution obtained from the linear relaxation. We consider a variant, called RANDOM k, where the process of choosing random lightpath routing is repeated for k times. The lightpath routing with the highest MCLC value is selected as the final output. The repetition increases the probability of obtaining a good solution. As we will see in the next section, randomized rounding provides a much more efficient way of obtaining a survivable lightpath routing without sacrificing the quality of solution. V. SIMULATION In this section, we discuss our simulation results for the algorithms introduced in Section IV. We first compare the lightpath routing algorithms by solving the ILP directly and by randomized rounding. Next, we compare the survivability performance among different formulations. Finally, we investigate the different lower bounds of MCLC, and their effects on the MCLC value of the lightpath routing when used as an optimization objective. ) ILP vs Randomized Rounding: In this experiment, we use the NSFNET (Figure 4) as the physical topology. The network is augmented to have connectivity 4, which makes it possible to study the performance of the algorithms where a higher MCLC value is possible. We generated 350 random logical topologies with connectivity at least 4, and size ranging from 6 to 2 nodes. Using the formulation MCF w with weight function w MinCut (s, t) introduced in Section IV-A as our benchmark, we compare the performance of RANDOM 0 against solving the ILP optimally. Table I compares the average running time between the algorithms ILP and RANDOM 0 on various logical topology size. All simulations are run on a Xeon E GHz workstation with 4GB of memory, using CPLEX to solve

9 RANDOM 0 MCLC Fig. 4. The augmented NSFNET. The dashed lines are the new links. Average MCLC the integer and linear programs. As the number of logical nodes increases, the running time for the integer program algorithm ILP increases tremendously. On the other hand, there is no observable growth in the average running time for the algorithm RANDOM 0, which is less than a minute. In fact, our simulation on larger networks shows that the algorithm ILP often fails to terminate within a day when the network size goes beyond 2 nodes. On the other hand, the algorithm RANDOM 0 for MCF w is able to terminate consistently within 2 hours for very large instances with a 00-node physical topology and 50-node logical topology. This shows that the randomized approach is a much more scalable solution to compute survivable lightpath routings. Logical Topology Size Average Running Time (seconds) ILP RANDOM TABLE I. Average Running Time of ILP and RANDOM 0 In Figure 5, the survivability performance of the randomized algorithm is compared with its ILP counterpart. Each data point in the figure is the MCLC average of 50 random instances with the given logical network size. As our result shows, the lightpath routings produced by RANDOM 0 have higher MCLC values than solving the ILP optimally. This is because the objective value for ILP MCF w is a lower bound on MCLC. As we will see in Section V-3, this lower bound is often not tight enough to accurately reflect the MCLC value, and the optimal solution to the ILP does not necessarily yield a lightpath routing with maximum MCLC. On the other hand, the randomized algorithm generates lightpath routings non-deterministically based on the optimal fractional solution of MCF w. Therefore, it approximates the lightpath routing given by the ILP, with an additional randomization component to explore better solutions. When the randomized rounding process is repeated many times, the algorithm often encounters a solution that is even better than the one given by the ILP. To sum up, randomized rounding provides an efficient alternative to solving integer programs without observable quality degradation. This allows us to experiment with more complex formulations in larger networks where solving the integer programs optimally is infeasible. In the next section, we will compare the different formulations introduced in Section IV-A, using randomized rounding to compute the lightpath routings. 2) Lightpath Routing with Different Formulations: In this experiment, we study the survivability performance of the Logical Topology Size Fig. 5. MCLC Performance of Randomized Rounding vs ILP lightpath routings generated by the formulations introduced in Section IV. We use the 24-node USIP network (Figure 6), augmented to have connectivity 4, as the physical topology. We generate 500 random graphs with connectivity 4 and size ranging from 6 to 5 nodes as logical topologies. Fig. 6. The augmented USIP network. The dashed lines are the new links. We compare the MCLC performance of the lightpath routings generated by the randomized rounding algorithm, RANDOM 00, on the following formulations: ) Multi-Commodity Flow MCF w, with the identity weight function, i.e., w(s, t) = for all (s, t) E L (Identity); 2) Multi-Commodity Flow MCF w, with the weight function w MinCut introduced in Section IV-A (MinCut); 3) Enhanced Multi-Commodity Flow MCF LF (LF). For comparison, we also run randomized rounding on the Survivable Lightpath Routing formulation (SURVIVE), introduced in [7], which computes the lightpath routing that minimizes the total fiber hops, subject to the constraint that the MCLC must be at least two. Figure 7 compares the average MCLC values of the lightpath routings computed by the four different algorithms. Overall, the formulations introduced in this paper achieve better survivability than SURVIVE. This is because these formulations try to maximize the MCLC in their objective functions, whereas SURVIVE minimizes the physical hops. Therefore, even though SURVIVE does well in finding a survivable routing (i.e. MCLC 2), the new formulations are able to achieve even higher MCLC values, which allow more physical failures to be tolerated. To further verify the survivability performance of the lightpath routings from a different perspective, for each lightpath routing, we simulated the scenario where each physical link fails independently with probability 0.0. Figure 8 shows the average probability that the logical topology becomes disconnected under this scenario. The

10 0 result is consistent with Figure 7, as lightpaths routings with higher MCLC values can tolerate more physical failures, and the logical topologies are thus more likely to stay connected. Average MCLC Average Probability Logical Topology Size LF MinCut Identity SURVIVE Fig. 7. MCLC Performance of Different Formulations LF MinCut Identity SURVIVE Logical Topology Size Fig. 8. Probabilty that Logical Topology Becomes Disconnected if Physical Links Fail Independently with Probability 0.0 The quality of the lightpath routing also depends on the graph structures captured by the formulations. Compared with MCF Identity, the formulation MCF MinCut uses a weight function that captures the connectivity structure of the logical topology. As a result, the algorithm will try to avoid putting edges that belong to smaller cuts onto the same physical link, thereby minimizing the impact of a physical link failure on these critical edges. This allows the algorithm MCF MinCut to produce lightpath routings with higher MCLC values than MCF Identity. As the number of logical links increases with the size of the logical topology, the maximum fiber link load becomes less effective in capturing the impact of fiber failures for the logical cuts. As a result, the performance of the simple multicommodity formulations degrades somewhat rapidly. On the other hand, the enhanced formulation MCF LF captures the connectivity structure of the logical topology in much greater detail, by having a constraint to describe the impact of a physical link failure to each logic cut. Therefore, this formulation is able to provide lightpath routings with the highest MCLC values. This observation is supported by Theorem 9, which states that the objective maximized by MCF LF is closer to the actual MCLC value. 3) Lower Bound Comparison: In Theorem 9 we establish different lower bounds for the MCLC. In this experiment, we measure these lower bound values for 500 different lightpath routings, and compare them to the actual MCLC values. As Figure 9 shows, the Weighted Load Factor is a very close approximation of the Min Cross Layer Cut. Among the 500 routings being investigated, the two metrics are identical in 368 cases. This suggests a tight connection between the two metrics, which also justifies the choice of such metrics as survivability measures. The figure also reveals a strong correlation between the MCLC performance and the tightness of the lower bounds given by the multi-commodity flow formulations in Section IV-A. Compared to MCF w, the formulation MCF LF provides an objective value that is closer to the actual MCLC value of the lightpath routing. This translates to better lightpath routings, as we saw in Figure 7. Since there is still a large gap between the MCF LF objective value and the MCLC value, this suggests room for further improvement with a formulation that gives a better MCLC lower bound. The summarize this section, a good formulation that properly captures the cross-layer connectivity structure is essential for generating lightpath routings with high survivability. Combined with randomized rounding, it gives a powerful tool for designing highly survivable layered networks. Metric Value Logical Topology Size MCLC WLF ILP LF ILP MinCut Fig. 9. Comparison among Min Cross Layer Cut (MCLC), Weighted Load Factor (WLF), and the Optimal Values of ILP LF and ILP MinCut VI. CONCLUSION In this paper, we introduce the problem of maximizing the connectivity of layered networks. We show that survivability metrics in multi-layer networks have significantly different meaning than their single-layer counterparts. We propose two survivability metrics, the Min Cross Layer Cut and the Weighted Load Factor, that measure the connectivity of a multi-layer network, and develop linear and integer formulations to compute these metrics. In addition, we use the metric Min Cross Layer Cut as the objective for the survivable lightpath routing problem, and develop multi-commodity flow formulations to approximate this objective. We show, through simulations, that our algorithms produce lightpath routings with significantly better Min Cross Layer Cut values than existing survivable lightpath routing algorithm. Our simulation shows that a good formulation, combined with the randomized rounding technique, provides a powerful tool for generating highly survivable layered networks. Therefore, an important direction for future research is to establish